Long Tail Keyword Suggestion for Sponsored Search Advertising

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 04 Issue: 07 | July -2017 www.irjet.net p-ISSN: 2395-0072
© 2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 1407
Long Tail Keyword Suggestion for Sponsored Search Advertising
Ms. Rinki Talwar1, Dr. Shuchita Upadhyaya2
1M.Tech Scholar, Dept. of Computer Science & Applications, Kurukshetra University, Haryana, India
2Professor, Dept. of Computer Science & Applications, Kurukshetra University, Haryana, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Sponsored Search Advertising is a method of
placing online advertisements along with organic results on
the Search Engine Result Page that will be displayed when a
user enters a query on search engine. Search advertising is
sold and delivered on the basis of keywords. The main
problem in Sponsored Search Advertising is of keyword
suggestion. In past, the advertisers tend to bid for the
keywords that have more search volume rather than that of
having low search volume. Hence, the bidding price of
former is more than that of latter. In this paper, an
improved topic modeling based approach is proposed to
suggest the related long tail keywords for advertisers that
have low volume and are inexpensive but generates the
equal amount of traffic cumulatively. Experimental results
on AOL search data, 2006 shows that the proposed
approach performs better than existing keyword suggestion
method.
Key Words: Long Tail Keywords, LightLDA, Sponsored
Search Advertising, Keyword Suggestion, AOL Search
Data.
1. INTRODUCTION
1.1 Background
Search engine advertising has nowadays become the
major evolving technology. There are two kinds of
advertisements: (a)Banner advertisements: It is a type of
display ads usually used by the large advertisers, and
(b)text advertisements or keyword advertisements or
Sponsored search advertisements: It is a type of textual
ads which are used mainly by all scale businesses and this
type of ads contains title and a short description of the
services offered by the advertisers, and also it contains the
URL which will direct the user to the advertiser’s website
[1]. Sponsored search ads are based on the queries
entered by the user online on search engine and deliver
many relevant results so, they are considered less
intrusive than that of banner advertisements or pop-up
advertisements. In this research paper, the focus is on the
text advertisements or Sponsored search advertisements.
Textual ads are the major part of Internet-Marketing.
While setting up a campaign with text ads, the advertisers
are asked to associate the keywords with their
advertisements to which it best describes. There are two
types of text advertisements system: (i) Keyword Targeted
Advertisement System, for ex. Google’s AdWords which
places the ads on the search result page [1] by matching
the keywords entered by the user with the keywords of
the advertiser’s ads, and (ii) Content Targeted
Advertisement System, for ex. Google’s AdSense [1], which
places the advertisements on the content-rich websites
such as newspapers etc. There exist numerous tools that
provide keyword suggestions to the advertisers for their
advertisement for ex. Google’s WordStream. There are
various sources on the basis of which keywords can be
extracted and suggested to the advertisers which are, (i)
Query log based keywords suggestion, (ii) Proximity based
keyword suggestion, and (iii) Meta-tag crawler-based
keyword suggestion.
The nature of keyword suggestions for the sponsored
search advertisements can be short-tail and long-tail.
Long-Tail keywords are keyword phrases made up of 3-5
or more words. The price of the head queries i.e. short-tail
queries are higher than that of the long tail queries b’coz
the competition between the advertisers for the short-tail
queries is higher than the long-tail queries. The long-tail
query suggestions for advertisements shows the best
results as long tail suggestions depicts the context of the
user [2] [3]. According to NEIL PATEL [4], the long query
suggestions for advertisers generate high traffic on their
advertisements rather than using Head Terms only (Fig-
1). That’s why, advertisers prefer to use the long keywords
because the customers who type descriptive keywords are
more qualified than those who type only short queries and
thus their advertisements reach to those customers
directly. The Example is shown in (Fig-2), which conveys
that the keyword ‘shoes’ generates HighCost and
Competition but, is not depicting the user’s intentions
exactly whereas the keyphrase ‘red Nike men’s Running
Shoes’ depicts the user intentions very clearly and have
Less Cost also.
1.2 Authors' Contribution
In this paper, the long-tail keyword suggestion system
for sponsored search advertisements is proposed. In [5]
Qiao proposed a keyword suggestion model for web
advertisement using the famous Topic Modeling, LDA with
Gibbs Sampling Approach, but its sampling complexity is
more. In this paper, the LightLDA an improved topic
modeling approach is used instead of using GibbsLDA [5],
because sampling complexity of LightLDA is less than that
of GibbsLDA as it uses the Metropolis-Hastings-Sampler
that constructed proposals very carefully and also results
in high convergence rate.
LightLDA also decreases the running time of the model
proposed in [5]Qiao work that uses GibbsLDA. Proposed

Work is targeted to increase the Precision, Recall, F1-
measure and decrease the running time.
The objectives of this paper are:
 To increase the accuracy of Qiao et al. Model [5].
 To decrease the running time of the Qiao et al.
Model [5].
The rest of this paper is organized as follows: in Sect.2 the
literature review related to the keyword suggestion
methods for the Sponsored Search Advertising is
presented. In Sect.3 Lda using gibbs sampling is described.
The proposed framework is discussed in Sect.4. The
experimental setup and experimental results are
described in Sect.5 & Sect.6 respectively and finally the
conclusion is outlined in Sect.7.
2. RELATED WORK
Various methods have been proposed for suggesting the
keywords for advertisements. According to various types
of data sources, the keyword suggestion methods are
broadly classified into 3 major categories:
 Keyword suggestion on basis of Query Log: In this
method, the keywords are suggested by using query
log of search engine. In the query log based method,
the keywords are suggested by conducting the co-
occurrence analysis in search engine query logs [6]
[7]. For example ‘Nexus’ is found to associate with
‘Google’ mainly, so the Nexus can be suggested as the
competitive keyword of the ‘Google’ (i.e. seed
keyword) to the advertisers. In [8], Zhang et al.
proposed a relevant but less competitive keyword
suggestions to the advertisers in order to boost the
revenue of the search engine and to fill the empty ad
slots. In [1] Sarmento et al. proposed the “synonymy”
method of suggesting keywords by mining the
previously submitted ads and thereby finding the
relevant and irrelevant keywords and thus suggesting
the relevant to the advertisers. In [9] DA et al. propose
a new method which depicts that query logs timely
measures the user’s intentions and can be used widely
in commercial advertising. In [10] Chuklin et al.
proposed the good query expansion methods that will
satisfy the user needs as the suggested query contains
the seed term. In [11]Chen et al. proposed the
“concept based suggestion method” in which the
keywords are suggested according to the conceptual
information extracted rather than the statistical
occurrence in the query log. In [12], Jiang et al.
proposed a framework on basis of which the queries
in the log that matches best with the input query
leading character is suggested to the users. In [13]
Szpektor et al. proposed a new method of long-tail
queries suggestions on the basis of Query-Flow-Graph
and concentrated on the query templates rather than
individual query-transitions. In [14], the author
proposed a ‘TermsNet’ method in which the semantic
relations between the terms and the neighbor of a
term is found out using the directed graph and then
the non-obvious words got rejected and important are
suggested to the advertisers. In [5] Qiao et al.
proposed a technique in which competitive keywords
are generated as output where in first step the
candidates are generated using indirect association
analysis on query log and after that the topic modeling
GibbsLDA approach is applied on the candidates set
which returns the competitive i.e. related keyword
suggestions to the seed keyword.
 Keyword suggestion on basis of Proximity with
seed: In this method, the keywords which are having
the high proximity with the seed keyword are
suggested. In [15] Abhishek and Hosanagar proposed
a method which suggests the keyword by finding the
semantic similarity between the terms by constructing
the similarity graph and suggesting the similar but
cheaper keywords. In [16]Broder et al. proposed a
method of classifying the user’s queries and then
suggesting the keywords similar to the entered
queries topics. In [11], some researches also uses the
thesaurus/dictionary (corpus already constructed by
the researcher) as a measure of proximity calculation.
There are various suggestion methods of this kind,
besides this it is not used much nowadays because it
doesn’t encounter the user real intentions which are
the primary concern of the advertisers.
 Keyword suggestion on basis of Meta-tag crawlers:
In this method, the keywords are extracted from the

meta-tags of the similar type of advertisements, for ex.
the seed keyword is entered to the search engine and
the meta-tags from the relevant web-advertisements
and web-pages got extracted and are suggested as
keywords to the advertisers. But this kind of methods
has some problems such as, firstly it doesn’t
concentrate on the advertiser’s concerns and secondly
advertisers may not be able to get many important
keywords.
3. LDA USING GIBBS SAMPLING
LDA (Latent Dirichlet Allocation) is a topic modeling
approach used to extract hidden patterns from the textual
documents. It is an unsupervised machine learning
technique. Latent dirichlet allocation is one type of topic
model and was firstly presented as a graphical model by
David Blei, Andrew Ng, Michael I. Jordan [17], and many
variations of this has come till now. The GibbsLDA i.e. LDA
using gibbs sampling assumes that each document is a
distribution over topics and further each topic is
distribution over words. LDA model is given in Fig-3. LDA
uses 2 dirichlet distributions θ and ∅ with two parameters
α and β respectively, denoted by θ~Dirichlet (θ|α) and
Ø~Dirichlet (Ø|β).
Fig-3 Plate Notation of LDA using gibbs sampling
Here, M is denoting the number of documents, N is
denoting the number of words in the document and lastly K
is denoting the number of topics. All, the other parameters
are described below:
α is the parameter of the dirichlet prior on the per-
document topic distributions,
β is the parameter of the dirichlet prior on the per-topic
word distribution,
M is the topic distribution for document M,
ØK is the word distribution for topic K,
ZM,N is the topic for the n-th word in document M, and
WM,N is the specific word.
Qiao et al. uses the GibbsLDA approach in their proposal
[5] which has certain limitations.
3.1 Limitations of LDA using Gibbs Sampling
There are certain limitations of LDA:
 Firstly, the speed of gibbs sampling inference method
is too slow for large dataset with many topics.
 Secondly, the sampling complexity of gibbs sampling
is more.
 Thirdly, the topics learned by LDA sometimes are
difficult to interpret by end users. (i.e. accuracy of LDA
using Gibbs sampling is not so good).
 Fourthly, LDA suffers from instability problem, which
occurs not only when there is new data arrives and
the model needs to be update but also when the same
Gibbs sampling method is run multiple times on the
same data.
To remove these limitations of LDA using gibbs
sampling, the LightLDA an improved topic modeling
approach is used in Qiao et al. model [5] instead of LDA
using gibbs sampling to enhance the performance of that
system.
4. PROPOSED FRAMEWORK
The proposed approach has been given the name
“LightLDA based keyword suggestion method” because it
uses the LightLDA topic modeling approach. It consists of
2 steps: (i) Candidate Set Generation, and (ii) Improved
Topic modeling.
The whole process is depicted in the model (Fig.-4).
Fig-4 Architecture of LightLDA based keyword suggestion
method
4.1 Candidate Set Generation
The very first step is to generate the candidate
keywords for the seed keyword using the query log. Query
Log used here is the AOL Query Log, 2006 which is
available online and consists of queries that user uses to
search online on AOL search engine. The clustered form of
AOL data is available from [18] which consists of 2
elements Query Keywords(i.e. represented by q.k) and its
volume, that is how many times a user uses this keyword

for searching(i.e. represented by q.k_vol). Sample data of
query log is shown in (Table-1).
Table -1: Query Log
The seed keyword, say‘s’ is the keyword regarding which
the long-tail keyword suggestions are to be generated for
the advertisers for their sponsored search advertisements.
The candidates are indirectly associated keywords to the
seed keywords. The method used to generate the
candidates is Candidate_Generation. Here, in this method
firstly the keywords directly associated with the seed is
found out which is denoted by s.AK and then the keywords
directly associated with s.AK and also indirectly associated
with the seed are found out which forms the candidate set
and is denoted by cCand. Corp denotes the corpus
consisting of directly and indirectly associated keywords.
For example, if a query ‘Apple iPad’ is taken then ‘iPad’ is
co-occurring keyword with the keyword ‘apple’, it means
keyword ‘iPad’ is associated with the keyword apple.
Candidate_Generation:
Input: Query-Log ‘Q’, Seed ‘S’
Output: Candidates Keyword Set, cCand(c1,c2,c3)
Begin:
1. Initialize arrays ‘s.AK’ , ‘corp’ , ‘cCand’ and ‘list’ to φ.
2. For each q in Q
3. if q contains S
4. add q to list
5. end
6. s.AK=s.AK U FindAssociativeKeyword(S,Q) //Method to
find the associative keyword
7. For each Query ‘q1’ in s.AK
8. cCand=cCand U FindAssociativeKeyword(q1,Q )
9. end
10. For each Query ‘q’ in s.AK
11. add q to corp
12.end
13. For each Query ‘q1’ in ccand
14. add q1 to corp
15.end
16. return corp, cCand and list
4.2 Improved Topic Modeling
This is the second step of the proposed method in
which the improved topic modeling approach ‘LightLDA’ is
applied on the candidate Keywords set ‘cCand’ generated
in first step and in the output of this step, the Long Tail
Keyword suggestions are retrieved for the sponsored
search advertisements.
Light LDA: It uses the cyclic Metropolis Hastings algorithm
combined with alias tables for both document-topic and
that of topic-word distribution. It uses the factorized
strategy of proposal instead of using a single proposal like
in GibbsLDA in [5]. LightLDA speeds up the process as
well as reduces the computational complexity. It divides
the proposal probability into 2 multiplicative terms
document proposal pd(k) and word proposal pw(k). To
construct the 2 proposals, document proposal as well as
the word proposal the true conditional probability of the
topic indicator zdi is given as:
where,
α and β are the 2 dirichlets used by standard GibbsLDA.
̅ = number of tokens in the document that are assigned to
the topic k,
nkw
-di = number of tokens with word ‘w’ that are assigned
to topic k, and
nk
–di = number of tokens assigned to topic k.
Main motive behind using LightLDA is to have high
proposal acceptance rates, good space coverage and
simplify the proposal generation complexity. For the
document-proposal, it generates the proxy alias table
which stores the number of times each topic appears in
the document and For word-proposal the alias tables are
generated for each word and algorithm cycles between
these two proposals and also perform the MH-Tests for
the acceptance and rejection of the proposals and alias
table is computed every time the word is used. The
acceptance probabilities of doc-proposal and word-
proposal give the proposal where as the current topic is
calculated with the help of LDA’s update equation. Each
keyword of cCand is tested to check with which
probability it is related with the Seed keyword using
LightLDA. The initial parameters set for the LightLDA
algorithm are α=0.1, β=0.01, k be the number of topics=2
and mh-step=2.
Pseudo Code of LightLDA
1. For each documentd d ε D
2. For each word x in d
3. compute proposal by alias ‘coinflip()’ method;
4. compute w matrix which stores the unique words
per document
5. compute k matrix which stores the topics per
document
6. decrement d,w,k count matrix
7. if proposal==0 //doc proposal
8. choose index= random(0,number of words);
9. calculate probability p=z[d][index]
Query Keyword(q.k) Query Volume(q.k_vol)
Cingular 7344
Horoscope 2500
Colgate 500
Google 1500
Wikipedia 200

10. calculate mh_acceptance=
compute_doc_Acceptance(k,p)
11. else //term proposal
12. p=alias_sample(w)
13 calculate mh_acceptance=
compute_term_acceptance(w,p)
14. end
15. end
16. // MH- test
17. mh_sample=random_float(0,1)
18. if mh_sample<mh_Acceptance
19. increment count matrix d,w,k //proposal is
rejected
20. revert to k
21. else
22. increment count matrix d,w,p //proposal is
accepted
23. end
24.end
By applying the LightLDA on candidate set cCand
generated in first step, the competitive keywords are find
out, the competitive keywords here defines are the
keywords that are related to the seed keyword and then
that competitive keywords are searched in the ‘list’
returned by the method ‘Candidate_Generation’ and those
queries that contains the competitive keywords are
returned as Long-Tail keyword suggestions.
5. EXPERIMENTAL SETUP
The implementation environment for performing
experiments was a Windows 7 system on Dell Pc with a
memory of 500GB & Intel core i5 processor and 4GB of
RAM (Random Access Memory). The project is created and
implemented in the java language using NetBeans IDE tool
of version 8.1.
5.1 Dataset
The dataset which is used here is obtained online from
AOL Query Log (2006) and consists of following fields
<AnonId, Query, Query Time, Item Rank, ClickURL>. The
clustered form of AOL data is available from [18] which
consists of 2 elements Query Keywords(i.e. denoted by q.k)
and its volume, that is how many times a user uses this
keyword for searching(i.e. denoted by q.k_vol) which is
shown in (Table-1).
Five seed keywords selected from different domains for
experiments is shown in Table-2.
Table-2 Seed Keyword
5.2 Evaluation Criteria
The relevancy criteria to evaluate the performance of
the proposed method used here is threshold, which is
described as:
If the volume of the query containing the seed
keyword in query log>2, then it is relevant suggestion.
5.3 Measures
Metrics used for evaluation are: Precision, Recall, F1-
measure, and running time which are discussed below one
by one.
 Precision: It is defined as the fraction of relevant
instances among the retrieved instances. It is also
known as the number of correct results retrieved
divided by the total number of retrieved results.
Mathematically can be calculated as,
 Recall: It is defined as the number of correct results
retrieved divided by the total number of relevant results
that should have been retrieved. Mathematically can be
calculated as,
 F1-measure: It is a measure of the total accuracy. It
depends on the Precision and Recall. Mathematically
can be calculated as,
 Running time: It is defined as the time required in
single execution of the proposed model. It is measured
in ms (milliseconds).
6. EXPERIMENTAL RESULTS
In this section, the comparative results between
proposed LightLDA based keyword suggestion method
and Qiao et al. proposed GibbsLDA based keyword
suggestion method [5] are shown below in Table-3
Table-3 Performance Comparison between GibbsLDA [5]
and LightLDA
Performance metrics GibbsLDA based
keyword suggestion
LightLDA based
keyword suggestion
Precision 0.85 0.883333
Recall 0.930657 0.967153
F1-Measure 0.888502 0.923345
Running Time(in ms) 790 391
Index Seed
1 Colgate
2 Budweiser
3 Sony
4 Pantene
5 Skype

Precision: In the proposed method, the precision got
increased as shown in Fig-5 also. It is more than that of the
GibbsLDA based keyword suggestion method of Qiao et al
[5]. Precision shows the average probability of relevant
retrieval. Increased values of precision indicate that the
more relevant results are retrieved among the suggested
results.
Fig-5 Result analysis of Precision
Recall: In the proposed method, the recall is more than
that of GibbsLDA based keyword suggestion method of
Qiao et al. [5] as shown in the Fig-6. Recall shows the
average probability of complete retrieval. Increased values
of recall will increase the F1-measure value directly.
Fig-6 Result analysis of Recall
F1-measure: It is the harmonic mean of precision and
recall. If both of them increase, F1-measure also get
increased. The comparative graph between proposed
LightLDA based keyword suggestion and GibbsLDA based
keyword suggestion of Qiao et al. is shown in Fig.-7.
Fig-7 Result analysis of F1-Measure
It is the true indicator of accuracy. Higher values of the F1-
measure of proposed method than that of GibbsLDA based
keyword suggestion method of Qiao et al. [5] indicates the
higher accuracy of keyword suggestion technique used in
proposed method.
Running Time: The running time of the proposed method
is less than that of the Qiao et al. proposed GibbsLDA
based keyword suggestion method in [5]which is shown in
the Fig.-8. The small running time of the proposed model
shows that the speed of the proposed method is higher
than the existing method [5].
Fig-8 Result analysis of Running Time
7. CONCLUSIONS
In this paper, the keyword suggestion system for
sponsored search advertisements has been proposed. The
objective is to suggest Long-Tail keywords to the
advertisers for sponsored search advertisements. We
designed the system by using improved topic modeling on
Query Log to suggest the Long-Tail keyword to
advertisers. The experimental results indicate that the
proposed method shows the better results than the topic
modeling GibbsLDA based approach proposed by Qiao et
al [5]. Also, the proposed system speeds up the process as
the running time it takes is less than that of the GibbsLDA
based approach proposed in [5].
REFERENCES
[1] Luís Sarmento, Paulo Trezentos, João Pedro
Gonçalves, and Eugénio Oliveira, "Inferring local
synonyms for improving keyword suggestion in an
on-line advertisement system," in ADKDD '09
Proceedings of the Third International Workshop on
Data Mining and Audience Intelligence for
Advertising, Paris, France, 2009, pp. 37-45.
[2] Amar Budhiraja and P. Krishna Reddy, "An Approach
to Cover More Advertisers in Adwords," in 2015 IEEE
International Conference on Data Science and
Advanced Analytics (DSAA), Paris, France, 2015.
[3] Amar Budhiraja and P. Krishna Reddy, "An Improved
Approach for Long Tail Advertising in Sponsored
0.82
0.84
0.86
0.88
0.9
Gibb'sLDA LightLDA
Precision
Precision
0.9
0.92
0.94
0.96
0.98
Gibb'sLDA LightLDA
Recall
Recall
0.86
0.88
0.9
0.92
0.94
Gibb'sLDA LightLDA
F1-measure
F1-measure
0
200
400
600
800
1000
Gibb'sLDA LightLDA
Running Time
Running Time

Search," in International Conference on Database
Systems for Advanced Applications, vol. 10178,
Suzhou, China, 2017, pp. 169-184.
[4] [Online]. http://neilpatel.com/blog/
[5] Dandan Qiao, Jin Zhang, Qiang Wei, and Guoqing
Chen, "Finding competitive keywords from query logs
to enhance search engine advertising," vol. 54, no. 4,
pp. 531-543, 2017.
[6] Wenhong Luo, David Cook, and Eric J. Karson, "Search
advertising placement strategy: Exploring the efficacy
of the conventional wisdom," vol. 48, no. 8, pp. 404-
411, 2011.
[7] Anton Schwaighofer, Joaquin Quiñonero Candela,
Thomas Borchert, Thore Graepel, and Ralf Herbrich,
"ADKDD '09 Proceedings of the Third International
Workshop on Data Mining and Audience Intelligence
for Advertising," in Knowledge Discovery and Data
Mining, Paris, France, 2009, pp. 27-36.
[8] Ying Zhang, Weinan Zhang, Bin Gao, Xiaojie Yuan, and
Tie-Yan Liu, "Bid keyword suggestion in sponsored
search based on competitiveness and relevance," vol.
50, no. 4, pp. 508-523, 2014.
[9] ZHI DA, JOSEPH ENGELBERG, and PENGJIE GAO, "In
Search of Attention," vol. LXVI, no. 5, pp. 1461-1499,
2011.
[10] Aleksandr Chuklin and Pavel Serdyukov, "How query
extensions reflect search result abandonments," in
SIGIR '12 Proceedings of the 35th international ACM
SIGIR conference on Research and development in
information retrieval, Portland, Oregon, USA, 2012,
pp. 1087-1088.
[11] Yifan Chen, Gui-Rong Xue, and Yong Yu, "Advertising
keyword suggestion based on concept hierarchy," in
WSDM '08 Proceedings of the 2008 International
Conference on Web Search and Data Mining, Palo
Alto, California, USA, 2008, pp. 251-260.
[12] Jyun-Yu Jiang, Yen-Yu Ke, Pao-Yu Chien, and Pu-Jen
Cheng, "Learning user reformulation behavior for
query auto-completion," in SIGIR '14 Proceedings of
the 37th international ACM SIGIR conference on
Research & development in information retrieval,
Gold Coast, Queensland, Australia, 2014, pp. 445-454.
[13] Idan Szpektor, Aristides Gionis, and Yoelle Maarek,
"Improving Recommendation for Long-tail Queries,"
in WWW '11 Proceedings of the 20th international
conference on World wide web, Hyderabad, India,
2011, pp. 47-56.
[14] Amruta Joshi and Rajeevi Motwan, "Keyword
Generation for Search Engine Advertising," in Sixth
IEEE International Conference on Data Mining -
Workshops (ICDMW'06), Hong Kong, China, 2006, pp.
490-496.
[15] Vibhanshu Abhishek and Kartik Hosanagar,
"Keyword generation for search engine advertising
using semantic similarity between terms," in
International Conference on E-Commerce,
Minneapolis, MN, USA, 2007, pp. 89-94.
[16] Andrei Z. Broder et al., "Robust classification of rare
queries using web knowledge," in SIGIR '07
Proceedings of the 30th annual international ACM
SIGIR conference on Research and development in
information retrieval, Amsterdam, The Netherlands,
2007, pp. 231-238.
[17] David M. Blei, Andrew Y. Ng, and Michael I. Jordan,
"Latent Dirichlet Allocation," vol. 3, pp. 993-1022,
2003.
[18] http://boston.lti.cs.cmu.edu/Data/web08-
bst/AOLQs.txt.

Long Tail Keyword Suggestion for Sponsored Search Advertising

Recommended

Recommended

More Related Content

Similar to Long Tail Keyword Suggestion for Sponsored Search Advertising

Similar to Long Tail Keyword Suggestion for Sponsored Search Advertising (20)

More from IRJET Journal

More from IRJET Journal (20)

Recently uploaded

Recently uploaded (20)

Long Tail Keyword Suggestion for Sponsored Search Advertising